Efficient Learning for Transferable Robot Autonomy in Human-Centered Environments

• Applicant: Professor Dr. Abhinav Valada
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 468878300

Tremendous progress in advancing the capabilities of autonomous robots in the last decade has enabled impressive results in the context of robot navigation in urban environments. These navigation abilities have been facilitated by obtaining directional instructions from pedestrians, leveraging highly accurate pre-build 2D or 3D maps, and even adapting their behavior using publicly available maps. However, most robots today still avoid being in the vicinity of humans and corresponding autonomy methods lack the generalization ability as well as the robustness required for ubiquitous service applications. Concurrently, data-driven machine learning algorithms have improved the performance of a multitude of robotic tasks. However, this has led to an increased dependency on manually annotated labels which are both environment and domain-specific, thereby restricting the transferability and generalizability of these methods. Moreover, navigation in dynamic environments shared with humans, such as on sidewalks still poses major challenges. This project aims to overcome these limitations by developing data-efficient and transferable learning techniques for fundamental navigational autonomy tasks to enable mobile service robots to navigate in thus far unknown urban outdoor environments among humans. The project addresses the following main questions:I) How can we obtain supervisory signals for learning models/policies without entirely relying onmanually annotated labels or reward functions?ii) How can we develop learning methods that can transfer learned knowledge across different tasks, modalities, and environments while being agnostic to robot hardware?iii) How can we learn navigation policies that generate adaptable and safe interaction behaviors that are socially compliant?The core objectives that we define are:Efficient Learning - Exploit self-supervisory signals from learning auxiliary tasks trained on unlabeled data, automatically generate labels from cross-modal supervision, exploit complementary cues and inductive transfer from simultaneously learning multiple tasks in a joint framework, instead of relying on manually annotated labels.Transferability - Develop learning techniques for multi-domain knowledge transfer, multi-domain fusion, and cross-domain adaptation, where the domain is either different modalities, robots, or conditions. Develop learning methods for actively seeking out informative states for adaptation and continuous learning without catastrophic forgetting.Human-Centered - Develop novel techniques to predict the motion and behavior of agents in human-centered environments in order to learn socially compliant and safe interaction behaviors for robot navigation.The outcome of this project will be an innovative approach to socially compliant urban navigation for mobile service robots that shows high independence of the sensor modalities and locomotion type, and that generalizes effectively to thus far unknown environments.

• DFG Programme: Independent Junior Research Groups
• Major Instrumentation: Spot Robot
• Instrumentation Group: 2320 Greif- und Hebewerkzeuge, Verladeeinrichtungen